[4/5] intel_iommu: Skip page walking on device iotlb invalidations

Commit Message

Eugenio Perez Martin Sept. 3, 2020, 4:14 p.m. UTC

Although they didn't reach the notifier because of the filtering in
memory_region_notify_iommu_one, the vt-d was still splitting huge
memory invalidations in chunks. Skipping it.

This improves performance in case of netperf with vhost-net:
* TCP_STREAM: From 1923.6Mbit/s to 2175.13Mbit/s (13%)
* TCP_RR: From 8464.73 trans/s to 8932.703333 trans/s (5.5%)
* UDP_RR: From 8562.08 trans/s to 9005.62/s (5.1%)
* UDP_STREAM: No change observed (insignificant 0.1% improvement)

Signed-off-by: Eugenio Pérez <[e-mail]>
---
 hw/i386/intel_iommu.c | 4 ++++
 1 file changed, 4 insertions(+)

Comments

Peter Xu Sept. 4, 2020, 6:32 p.m. UTC | #1

On Thu, Sep 03, 2020 at 06:14:45PM +0200, Eugenio Pérez wrote:
> Although they didn't reach the notifier because of the filtering in
> memory_region_notify_iommu_one, the vt-d was still splitting huge
> memory invalidations in chunks. Skipping it.
> 
> This improves performance in case of netperf with vhost-net:
> * TCP_STREAM: From 1923.6Mbit/s to 2175.13Mbit/s (13%)
> * TCP_RR: From 8464.73 trans/s to 8932.703333 trans/s (5.5%)
> * UDP_RR: From 8562.08 trans/s to 9005.62/s (5.1%)
> * UDP_STREAM: No change observed (insignificant 0.1% improvement)
> 
> Signed-off-by: Eugenio Pérez <[e-mail]>

Reviewed-by: Peter Xu <[e-mail]>

Patch

diff --git a/hw/i386/intel_iommu.c b/hw/i386/intel_iommu.c
index ab6833d5a0..fbbda0c87e 100644
--- a/hw/i386/intel_iommu.c
+++ b/hw/i386/intel_iommu.c
@@ -1478,6 +1478,10 @@  static int vtd_sync_shadow_page_table(VTDAddressSpace *vtd_as)
     VTDContextEntry ce;
     IOMMUNotifier *n;
 
+    if (!(vtd_as->iommu.iommu_notify_flags & IOMMU_NOTIFIER_IOTLB_EVENTS)) {
+        return 0;
+    }
+
     ret = vtd_dev_to_context_entry(vtd_as->iommu_state,
                                    pci_bus_num(vtd_as->bus),
                                    vtd_as->devfn, &ce);